Google has announced Gemini 4 Carbon, its latest artificial intelligence (AI) model, with improved coding capabilities, advanced cybersecurity features and the ability to handle complex tasks that may take several days to complete.
The company said the model ranked first on the Artificial Analysis Intelligence Index, with a score of 59. It also recorded improvements in software development, reasoning and tasks that require AI agents to work through multiple steps with limited human intervention.
Gemini 4 Carbon builds on Gemini 4 Argon, which Google launched two weeks ago. The company said its teams had tested Carbon on complex engineering problems, including identifying the root causes of bugs and fixing them across large software systems.
Focus on cybersecurity
A key feature of the new model is its ability to identify software vulnerabilities, assess whether they can be exploited by attackers and develop fixes.
Gemini 4 Carbon scored 88.2% on SEC-bench Pro, a benchmark that tests the ability to find security flaws in real-world software. It scored 54.8% on ExploitGym, which evaluates vulnerability exploitation, and 71% on CWE-bench, which tests the ability to fix software weaknesses.
Google said the model achieved more than twice the success rate of Gemini 4 Argon in generating secure code in internal tests. Its Chrome team also reported significantly better performance in generating software patches compared with another advanced AI model.
Improved accuracy and safety
Google said it had worked to reduce incorrect claims and improve the model’s ability to follow instructions without exceeding the limits set by users.
According to the company, unsupported claims fell by about half compared with Argon, while high-severity misalignment alerts in internal coding environments dropped by nearly 80%. The model also recorded a zero attack success rate in the evaluated indirect prompt-injection benchmark, which tests whether malicious instructions hidden in external content can manipulate an AI system.
Limited rollout initially
Google will initially make Gemini 4 Carbon available to a selected group of cybersecurity professionals before gradually expanding access to trusted partners. Wider availability for developers, businesses and consumers is planned later.
The introductory price is $2 per million input tokens and $10 per million output tokens. Cached input tokens will be available at a 95% discount on the input-token price.